Submission 235
CRUX
Submitter group: Individuals.
Position labels are not shown for individual submitters.
An AI model identified the submitter group and wrote the summary, key points and recommendation summaries. Counts do not measure importance or merit.
Summary
Read the original submission for the full text.
An independent researcher, who founded CRUX (a project developing evidence-bound AI assurance methods), argues AI assurance should be an evidence problem, with consequential AI claims linked to versioned, replayable evidence and reassessed when systems change. Recommends common vocabulary, agency assurance records, vendor disclosure, independent benchmark infrastructure and agent evaluations; notes an export opportunity in assurance tooling.
Key points
Quotes are the submitter’s own words. Text taken from documents, including scanned pages read by text-recognition software, can differ from the PDF. Check the original. Listed in the same order as on the submission’s page in the interactive guide.
Argues that AI assurance should be treated as an evidence problem rather than merely a disclosure or principles problem.
Australia should treat AI assurance as an evidence problem, not merely a disclosure or principles problem.
Argues that a test that did not observe a failure does not show a system cannot fail, and that a benchmark score or valid technical result does not establish safety or authorise deployment.
A test that did not observe a failure does not establish that the system cannot fail. A benchmark score does not establish production safety. A valid technical result does not itself authorise deployment.
Recommends that material changes to models, prompts, tools or data trigger selective reassessment of the claims that depend on them.
Material changes to a model, system prompt, tool configuration, data source, policy threshold, interface or deployment environment should trigger selective reassessment of the claims that depend on the changed component.
Recommends that technical evidence and organisational authority to approve be recorded separately.
Technical evidence and organisational authority should be recorded separately. A reviewer may establish that a test passed without having authority to approve deployment.
Recommends that agent evaluations inspect the full action pathway, including permissions and external writes.
Agent evaluations should inspect the full action pathway: intent, emitted call, serialization, parsing, execution, returned observation, repair or rescue, permission state, external writes and persistent state changes.
Argues that Australia can build a comparative advantage in AI assurance, evaluation and governance tooling.
Australia can build a comparative advantage in AI assurance, evaluation and governance tooling. This is a credible export and professional-services opportunity
Recommendations
Recommendations are shown beside the submission’s own words. Listed in the same order as on its page in the interactive guide.
For high-impact or government AI uses, record material claims with the proposition, evidence relied upon, provenance, version, scope, assumptions, missing evidence, dependencies, expiry or review conditions and the maximum conclusion the evidence licenses.
For high-impact or government AI uses, material claims should be recorded with: the exact proposition; evidence relied upon; source identity and provenance; model/system version; evaluation scope; assumptions; known missing evidence; dependencies; expiry or review conditions; and the maximum conclusion licensed by that evidence.
Have Australian assurance guidance distinguish 'not observed', 'not demonstrated' and 'cannot occur', requiring evidence that the evaluation path could observe the event for negative capability or safety claims.
Australian assurance guidance should explicitly distinguish 'not observed', 'not demonstrated' and 'cannot occur'. Negative capability or safety claims should require evidence that the relevant evaluation path was capable of observing the event.
Bind assurance to versions and selectively reassess dependent claims after material model, prompt, tool, data, policy-threshold, interface or environment changes, rather than rerunning every test.
Assurance should be version-bound. Material changes to a model, system prompt, tool configuration, data source, policy threshold, interface or deployment environment should trigger selective reassessment of the claims that depend on the changed component. The goal is not to rerun every test after every change, but to make dependencies explicit enough to know what must be rechecked.
Record technical evidence separately from organisational approval authority.
Technical evidence and organisational authority should be recorded separately.
Evaluate agents’ full action pathway: intent, calls, serialization, parsing, execution, observations, repair/recovery, permissions, external writes and persistent state.
Agent evaluations should inspect the full action pathway: intent, emitted call, serialization, parsing, execution, returned observation, repair or rescue, permission state, external writes and persistent state changes.
Where evaluation environments are accidentally too permissive, record the outer-system failure separately from model behaviour rather than collapsing both into one score.
Where an evaluation environment is accidentally more permissive than intended, that outer- system failure should be recorded separately from the model's behaviour rather than collapsed into a single score.
Ask the Australian AI Safety Institute and relevant regulators to publish a common evidence-to-claim vocabulary for consequential AI assurance.
Ask the Australian AI Safety Institute and relevant regulators to publish a common evidence-to-claim vocabulary for consequential AI assurance.
Require agencies procuring or deploying higher-impact AI to maintain versioned, replayable assurance records for material claims and approvals.
Require agencies procuring or deploying higher-impact AI to maintain versioned, replayable assurance records for material claims and approvals.
Require vendors to disclose which evaluation results depend on hidden judges, shared upstream evidence, synthetic data or incomplete observation paths.
Require vendors to disclose which evaluation results depend on hidden judges, shared upstream evidence, synthetic data or incomplete observation paths.
Adopt explicit change-trigger rules so that model, version or configuration changes selectively invalidate or reopen affected approvals.
Adopt explicit change-trigger rules so that model/version/configuration changes selectively invalidate or reopen affected approvals.
Fund independent, reproducible benchmark and assurance infrastructure that can compare frontier systems over time without depending on vendor self-attestation.
Fund independent, reproducible benchmark and assurance infrastructure that can compare frontier systems over time without depending on vendor self-attestation.
Encourage red-team and agent evaluations that test permission boundaries, external-state creation, restart/recovery and evidence-trace completeness.
Encourage red-team and agent evaluations that test permission boundaries, external-state creation, restart/recovery and evidence-trace completeness.
Use government procurement to reward transparent, reproducible assurance methods rather than proprietary confidence scores.
Government procurement can help create this capability by rewarding transparent, reproducible assurance methods rather than proprietary confidence scores.